Add mig support in specs

aws/infrastructure.tf

@@ -207,6 +207,7 @@ locals {
         cpus = data.aws_ec2_instance_type.instance_type[values.prefix].default_vcpus
         ram  = data.aws_ec2_instance_type.instance_type[values.prefix].memory_size
         gpus = try(one(data.aws_ec2_instance_type.instance_type[values.prefix].gpus).count, 0)
+        mig  = lookup(values, "mig", null)
       }
     }
   }

azure/infrastructure.tf

@@ -170,6 +170,7 @@ locals {
         cpus = local.vmsizes[values.type].vcpus
         ram  = local.vmsizes[values.type].ram
         gpus = local.vmsizes[values.type].gpus
+        mig  = lookup(values, "mig", null)
       }
     }
   }

common/configuration/main.tf

@@ -125,6 +125,7 @@ locals {
             public  = tls_private_key.ed25519[values.prefix].public_key_openssh
           }
         }
+        mig = lookup(values.specs, "mig", null)
       }
     )
   }

common/configuration/puppet.yaml

@@ -122,6 +122,12 @@ runcmd:
   - "(tar xf aws-efa-installer-latest.tar.gz && cd aws-efa-installer && ./efa_installer.sh --yes --minimal)"
   - rm -fr aws-efa-installer aws-efa-installer-latest.tar.gz
 %{ endif }
+%{ if mig != null }
+# Install nvidia-mig-manager to enable MIG without NVIDIA drivers installed
+  - yum -y install https://github.com/NVIDIA/mig-parted/releases/download/v0.5.5/nvidia-mig-manager-0.5.5-1.x86_64.rpm
+# It does not matter which config is selected, the goal here is to enable MIG before reboot
+  - nvidia-mig-parted apply -f /etc/nvidia-mig-manager/config.yaml -c all-1g.5gb --mode-only
+%{ endif }
 %{ if cloud_provider == "gcp" }
 # Google Cloud user-data fact generates a warning because its size is greater than what is allowed (<4096 bytes).
 # We have no use for it, so we remove startup-script, user-data and user-data-encoding when running in GCE.

docs/README.md

@@ -546,19 +546,23 @@ Optional attributes can be defined:
 2. `image`: specification of the image to use for this instance type. (default: global [`image`](#46-image) value).
 Refer to section [10.12 - Create a compute node image](#1012-create-a-compute-node-image) to learn how this attribute can
 be leveraged to accelerate compute node configuration.
-3. `disk_size`: size in gibibytes (GiB) of the instance's root disk containing
+3. `disk_type`: type of the instance's root disk (default: see the next table).
+    | Provider | `disk_type` | `disk_size` (GiB) |
+    | -------- | :---------- | ----------------: |
+    | Azure    |`Premium_LRS`| 30                |
+    | AWS      | `gp2`       | 10                |
+    | GCP      | `pd-ssd`    | 20                |
+    | OpenStack| `null`      | 10                |
+    | OVH      | `null`      | 10                |
+4. `disk_size`: size in gibibytes (GiB) of the instance's root disk containing
 the operating system and service software
-(default: see the next table).
-4. `disk_type`: type of the instance's root disk (default: see the next table).
-
-Default root disk's attribute value per provider:
-| Provider | `disk_type` | `disk_size` (GiB) |
-| -------- | :---------- | ----------------: |
-| Azure    |`Premium_LRS`| 30                |
-| AWS      | `gp2`       | 10                |
-| GCP      | `pd-ssd`    | 20                |
-| OpenStack| `null`      | 10                |
-| OVH      | `null`      | 10                |
+(default: see the previous table).
+5. `mig`: hash map of [NVIDIA Multi-Instance GPU (MIG)](https://docs.nvidia.com/datacenter/tesla/mig-user-guide/index.html) short profile names and count used to partition the instances' GPU, example for an A100:
+    ```
+    mig = { 1g.5gb = 2, 2g.10gb = 1, 3g.20gb = 1 }
+    ```
+    This is only functional when the [GPU is supported](https://docs.nvidia.com/datacenter/tesla/mig-user-guide/index.html#supported-gpus),
+    and with x86-64 processors (see [NVIDIA/mig-parted issue #30](https://github.com/NVIDIA/mig-parted/issues/30)).
 
 For some cloud providers, it possible to define additional attributes.
 The following sections present the available attributes per provider.

gcp/infrastructure.tf

@@ -180,6 +180,7 @@ locals {
         cpus = data.external.machine_type[values["prefix"]].result["vcpus"]
         ram  = data.external.machine_type[values["prefix"]].result["ram"]
         gpus = try(data.external.machine_type[values["prefix"]].result["gpus"], lookup(values, "gpu_count", 0))
+        mig  = lookup(values, "mig", null)
       }
     }
   }

openstack/infrastructure.tf

@@ -137,6 +137,7 @@ locals {
           parseint(lookup(data.openstack_compute_flavor_v2.flavors[values.prefix].extra_specs, "resources:VGPU", "0"), 10),
           parseint(split(":", lookup(data.openstack_compute_flavor_v2.flavors[values.prefix].extra_specs, "pci_passthrough:alias", "gpu:0"))[1], 10)
         ])
+        mig  = lookup(values, "mig", null)
       }
     }
   }